InSAR And Deep Learning Based Landslide Deformation Detection And Hidden Danger Identification

Due to its complex geological environment and geomorphological conditions,Guizhou Province,which is located in the mountainous region of southwest China,has frequent geological disasters under the impetus and human activities.Landslides are the main geological disasters.Landslide geological disasters have suddenness,complexity,high risk characteristic and frequent seasonal occurrence,often causing huge losses and even devastating disasters to people’s lives.It is difficult to achieve continuous and efficient monitoring of individual landslide on a large scale by the traditional method which need to arriving at the one of landslide site.This thesis takes the serious landslides disasters area in the northern part of Huishui County,Qi’nan Buyi Miao Autonomous Prefecture,Guizhou Province as the research area which is dominated by Lianjiang alluvial plain,and uses the time-series InSAR technology to identify deformation hazard areas.After that the study identifies potential landslide hazards from optical remote sensing visual interpreting of deformation hazard areas.Combined with deep learning methods for automatic identification of landslide and non-landslide targets for landslide hazards,and the research of landslide hazards based on the identification results.The main achievements have the following aspects:(1)It is considered that the choice of time baseline has great influence on the coherence and interference quality by exploring the effects of time baseline,spatial baselines and four typical land types: bare land,residential land and facilities,water,and vegetation on the coherence of SAR image pairs.Aiming at the problem of difficult extraction of deformation information in mountainous areas with high vegetation cover and large topographic relief,this thesis sets with time baseline of 36 days,60 days and 90 days were established to investigate the influence of time baseline on the deformation finally obtained.SBAS-InSAR technology was used to process the historical image data of 59 Sentinel-1A radar satellite from 2019 to2020 to get annual average deformation rate.The results show that the most accurate deformation information is extracted from the 60-day time baseline set in the study area,and there are many of the deformation information,fewer misjudgments and omissions than other time baseline sets.(2)The settlement amount and settlement rate of PS points in the study area were obtained by PS-InSAR technology,and the average deformation rate extracted by SBAS-InSAR based on 60 d time baseline set was comprehensively analyzed.It was found that the monitoring results of the two time-series InSAR technologies has consistency on the whole.The results show that the main subsidence areas are located in the central and eastern direction of the town of Yashui and the town of Pendleton with a maximum subsidence rate of-83.18mm/y.It also shows that the urban area of Mengjiang District,Lianjiang District and Haohuahong Town,the main towns on the Lianjiang alluvial plain,is relatively stable,with deformation rate less than 10mm/y.(3)15 deformation risk areas were delineated as potential landslide danger areas based on the surface deformation information extracted by InSAR technology,combined with optical remote sensing visual interpretation.A sample set of typical water-related landslides is created by using optical remote sensing images covering five typical landslide areas in Guizhou Province.(4)Using optical remote sensing images covering five typical landslide areas in Guizhou Province to create a sample set of typical water-bearing landslides in Guizhou Province.The sample set is using to train a convolutional neural network which improves by adding residual units.This Resnet’s training set has 72.89% accuracy and a Kappa coefficient of 0.6381The above results prove that the method combining InSAR technology with deep learning has certain feasibility for landslide target identification.It provides the idea of realizing landslide disaster identification in large mountainous area,completion of landslide hazard warning and prevention and improved effective monitoring system for potential landslide hazards in long time.